Pith. sign in

REVIEW 4 major objections 5 minor 57 references

IAR2: Improving Autoregressive Visual Generation with Semantic-Detail Associated Token Prediction

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read IAR2's central claim is that splitting each image patch into a semantic token and a residual detail token resolves the codebook-size trade-off in autoregressive image generation, yielding FID 1.50 on ImageNet 256×256.

desk verdict The semantic-detail dual-codebook idea is real and the ablations are honest, but the headline FID 1.50 is a 384×384 result misrepresented as 256×256, so the SOTA claim does not survive as written. read the letter →

arxiv 2510.06928 v2 pith:IGZOZFD7 submitted 2025-10-08 cs.CV

Ran Yi , Teng Hu , Zihan Su , Jiangning Zhang , Lizhuang Ma This is my paper
classification cs.CV
keywords autoregressiveimagegenerationsemantic-detaildualcodebookresidualvectorquantizationhierarchicalpredictionvisualtokenizationclassifier-freeguidanceNetlocalcontextmodeling
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Autoregressive image generation faces a tension: a larger visual codebook reconstructs images better, but it makes next-token prediction harder, and generation quality degrades once the codebook grows too large. IAR2 resolves this by giving each patch two tokens: a compact semantic code that captures global content and a larger detail code that captures residual texture. Generation then predicts the semantic token first and the detail token conditioned on it, expanding effective codebook capacity from n to n1×n2 while keeping each prediction step small. The paper reports that this design, together with a local-context prediction head and an adaptive guidance schedule, reaches FID 1.50 on ImageNet 256×256 with a 1.5B-parameter model, surpassing larger models trained with more GPUs. A sympathetic reader would care because it suggests that structured coarse-to-fine tokenization, not just scale, is what makes autoregressive visual generation competitive.

What carries the argument

The Semantic-Detail Associated Dual Codebook: a 256-entry semantic codebook and a 4096-entry detail codebook, trained by residual quantization (nearest semantic code first, then residual detail code), with semantic pretraining under a perceptual loss followed by joint training with reconstruction, adversarial, and commitment losses. The companion Local-Context Enhanced Autoregressive Head fuses the two token embeddings into one hidden state, predicts the semantic token, then conditions the detail token on it, using a compressed local-context window to enrich the prediction. The load-bearing identity is the factorization p(patch) ≈ p(semantic) × p(detail | semantic), which turns a one-step pr

What would settle it

On a held-out set, compute the per-token detail cross-entropy under three conditions: conditioned on the model's own sampled semantic token, conditioned on the ground-truth semantic token from the tokenizer, and unconditioned. If the first two differ strongly, the detail head has learned to rely on correct semantic input and free-running generation will carry avoidable error; if the third is close to the first, the semantic token carries little usable information for details, contradicting the factorization.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the reconstruction-versus-generation trade-off of a single codebook is not fundamental. By training a residual dual codebook — a compact semantic codebook for global content and a larger detail codebook for residual texture — the model obtains roughly a million effective discrete codes (256×4096) while the autoregressive model only ever predicts from the smaller semantic vocabulary and then from the detail vocabulary. The generative process factors each patch's prediction as p(semantic) × p(detail | semantic), with the detail token conditioned on the just-predicted semantic token, all within the same sequence position. The paper argues that this hierarch

Load-bearing premise

The load-bearing premise is that an image patch's visual content is well captured by an independent semantic code plus a residual detail code, so that p(patch) approximately factors as p(semantic) × p(detail | semantic); if details are mostly determined by neighboring patches rather than by the patch's own semantic label, or if training and inference condition the detail head differently, the reported gains could fail to transfer.

Editorial extensions

If this is right

  • A dual-codebook representation gives near-single-codebook reconstruction quality while keeping each prediction step small; reconstruction rFID drops to 1.05 with only 4352 total code entries.
  • Predicting semantic tokens first makes generation robust to detail-token errors, since the coarse content is fixed before fine texture is added.
  • Local-context enhancement in the prediction head improves spatial coherence and training efficiency, and the gains persist at higher-resolution 24×24 token grids.
  • Per-token adaptive CFG improves conditional alignment and diversity over any fixed guidance scale, by lowering guidance in background regions and strengthening it for condition-relevant regions over time.
  • The improvements hold across model sizes from roughly 100M to 1.5B parameters and across 50- and 300-epoch training schedules, indicating the architectural changes, not longer training, drive the gain.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The semantic-detail hierarchy suggests a general recipe for discrete visual tokenizers: pair a small, semantically meaningful codebook with a large residual codebook. This could extend to video or multi-view generation, where the semantic code carries object or scene identity and the detail code carries per-frame texture.
  • A testable extension is whether the detail codebook can be predicted more cheaply from neighboring patches' detail codes than from the patch's own semantic token; if so, a dedicated spatial detail predictor could reduce computation further or improve long-range texture consistency.
  • Because PAG-CFG computes per-token attention to the condition, the same attention maps could serve as an interpretability signal, revealing where the model relies on the class or text condition during generation.
  • The reported efficiency gain comes partly from keeping the sequence length unchanged; other autoregressive image models could likely adopt the dual-codebook retokenization and hierarchical prediction head without changing their backbone architecture.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes IAR2, an autoregressive image generation framework built around a Semantic-Detail Associated Dual Codebook, a local-context-enhanced autoregressive head, and a progressive attention-guided CFG mechanism. The method is evaluated on class-conditional ImageNet generation across model sizes from 143M to 1.5B parameters, with extensive ablations of codebook size, prediction paradigm, AR-head compression, codebook association, and CFG hyperparameters. The central claim is that the hierarchical semantic-then-detail prediction and the dual-codebook representation improve generation fidelity and efficiency, culminating in a reported state-of-the-art FID of 1.50 on ImageNet 256×256.

Significance. The core idea of decoupling visual tokens into semantic and detail codebooks and predicting them hierarchically is well-motivated and, if the experimental results are valid, would be a genuine advance for autoregressive visual generation. The paper is commendably thorough in its ablations: Tables 4–8 and the token-prediction comparisons in §§5.6–5.10 cover the main design choices, and the authors release code. However, the headline state-of-the-art claim is not currently supported by the reported data because (i) the top results in Table 2 appear to be for 384×384 images, not the 256×256 benchmark stated in the table header and abstract, and (ii) the best FID is obtained with rejection sampling that is not disclosed in the abstract. These are load-bearing issues for the claim, but they are fixable with additional experiments or a repositioning of the claims.

major comments (4)
  1. [Abstract & Table 2 vs. Table 3] The abstract and Table 2 claim state-of-the-art FID 1.50 on ImageNet 256×256, but Table 3 shows that the IAR2-XL/XXL rows in Table 2 (FID 2.10/IS 286.4, FID 1.76/IS 279.5, and the rejection-sampled FID 1.50) are exactly the 24×24-token (384×384) 300-epoch results. The 16×16-token (256×256) entries for XL/XXL are only trained for 50 epochs, where IAR2-XXL has FID 2.30. Therefore the published 1.50 result is not on 256×256, and comparing it with VAR-d30's 256×256 FID of 1.92 is an apples-to-oranges comparison. At 256×256, the best reported IAR2 result (2.30) does not beat VAR-d30. This inconsistency invalidates the headline SOTA claim as presented.
  2. [Table 2 and §5.3] The FID 1.50 is achieved only with rejection sampling (Table 2, IAR2-XXL*), but the abstract and §5.3 present it without qualification. The comparison with VAR-d30 uses VAR-d30's standard FID 1.92, which is not marked as rejection-sampled. Thus the advantage may be partially attributable to extra decoding-time compute rather than the proposed model. Please report the non-rejection FID at the same resolution and explicitly state that 1.50 uses rejection sampling.
  3. [§4.3.2, Eq. (13)-(14)] The hierarchical prediction scheme conditions detail prediction on the previously predicted semantic token k_{i+1}. It is not specified whether the training loss uses ground-truth semantic tokens (teacher forcing) or sampled/predicted tokens. If ground-truth tokens are used, then at inference the model conditions on potentially erroneous semantic predictions, creating a train/test mismatch. If predicted tokens are used, the gradient flow through the sampled token is not described. This is central to the factorization in Eq. (13) and affects reproducibility.
  4. [§5.3] The text states 'IAR2 achieves state-of-the-art performance, reaching an FID of 1.50 and an IS of 286.4.' However, IS 286.4 belongs to IAR2-XL (whose FID is 2.10), not to the IAR2-XXL* row with FID 1.50 (IS 282.7). This conflation of rows obscures the trade-off between model size and metrics and should be corrected.
minor comments (5)
  1. [§4.4] The local window size k is never specified in the main text or in the appendix hyperparameter tables. Please report the value used for the k×k local context.
  2. [§4.2] The default codebook sizes (256 semantic, 4096 detail) appear only in Table 1 and the appendix; the main text should state the default configuration explicitly when introducing the dual codebook.
  3. [§3.2] Equation numbering is inconsistent: Eq. (1) is followed directly by Eq. (4) in §3.3, and the cluster-level loss is labeled with (2) and (3) only after (4). Please renumber.
  4. [Table 2] The table header says 'class-conditional ImageNet 256×256' but the IAR2-XL/XXL rows and the corresponding LlamaGen/IAR rows are drawn from 24×24-token (384×384) settings. The table should either be restricted to 256×256 or clearly annotated by resolution.
  5. [§5.5 and Fig. 7] Fig. 7(a) reports 'horizontal lines' for PAG-CFG, but the figure as described does not show the actual PAG-CFG values on the same axes; please make the comparison quantitative in the figure.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: IAR2's central claims rest on measured FID/IS results and internal ablations; self-citations to IAR are motivational and used as baselines, not as load-bearing proof steps.

full rationale

The paper's central results are empirical: FID 1.50, IS 286.4, and comparisons with LlamaGen, IAR, and VAR are measured generation metrics, not quantities derived from model definitions. The claimed capacity increase from a single codebook to n1×n2 is a straightforward combinatorial description of the two-codebook construction (Eq. 5-6), not a prediction forced by a fit. The prior work IAR [9] is cited for inspiration and as a baseline, but the load-bearing evidence for IAR2's components is provided within this paper: Table 5 compares hierarchical prediction against independent and alternating prediction; Table 7 compares associated dual codebooks against single and unassociated dual codebooks; Table 8 isolates PAG-CFG. None of these reduce by construction to the inputs, and no fitted parameter is renamed as a prediction. Self-citation appears in the motivation ('our prior work, IAR, initiated a direction'), but the paper also includes its own embedding-similarity analysis in §3.1, so the cited prior work is not the sole support. The notable Table 2/Table 3 discrepancy (the reported 1.50 FID corresponds to the 24×24-token/384×384 setting while being presented under the 'ImageNet 256×256' benchmark) is a reporting-consistency or correctness concern, not a circularity concern: even if the headline number is overstated, it is still a measured number rather than a result whose derivation is equivalent to its inputs. No definitional circularity, fitted-input-as-prediction, or load-bearing self-citation chain was found.

Assumptions & free parameters 7 free parameters · 5 assumptions · 1 invented entities

The central claim rests on several fitted hyperparameters (codebook sizes, λ_s, CFG schedule, interleaving ratio) and on domain assumptions about image semantics and attention interpretability. Codebook training losses are incompletely specified. No new physical entities are introduced; the dual codebook is an architectural invention supported only by in-paper ablations.

free parameters (7)
  • semantic codebook size n1 = 256
    Chosen via ablation Table 1; sets granularity of coarse semantic prediction.
  • detail codebook size n2 = 4096
    Chosen via ablation Table 1; larger sizes (8192) hurt gFID.
  • semantic loss weight λ_s = 2.0 (B), 1.0 (L/XL/XXL), 1.5 (XXL 24×24)
    Tuned via ablation Table 10; balances semantic vs detail cross-entropy.
  • CFG start/end scales = e.g., 1.75→3.0 for IAR2-B; varies per model in Table 12
    Selected by grid search on FID (Table 9); the headline number depends on these values.
  • VQ interleaving ratio (2:1) = 2:1
    Chosen by hand in §4.2.2 to preserve semantic codebook quality.
  • local window size k = not disclosed
    Local-context window k×k in §4.4 affects the AR head, but its value is never given.
  • loss weights β, λ_perc, λ_rec, λ_adv = not disclosed
    Weights in Eq. 7–9 are not reported; they shape the learned codebooks.
assumptions (5)
  • domain assumption Visual embedding similarity implies visual/semantic similarity (Fig. 3, Sec. 3.1)
    Borrowed from IAR; verified on VQGAN, assumed to hold for the new semantic codebook.
  • domain assumption LPIPS loss is a suitable proxy for semantic content
    Stage-1 codebook training (Eq. 7) uses only LPIPS + commitment to define 'semantic'.
  • ad hoc to paper The sum-set of semantic and detail codebooks has effective size n1×n2
    Sections 4.2 and 4.3 claim polynomial capacity; this is an upper bound and assumes no harmful collisions in a fixed-dimension latent space, with no analysis of codebook geometry.
  • domain assumption Attention scores from image tokens to the conditioning token are a valid spatial relevance map
    PAG-CFG (§4.5) uses aggregated attention as the per-token multiplier α_i; no experiment verifies attention correlates with foreground/background.
  • domain assumption The next-token AR paradigm extends to two-codebook pairs
    The AR model is trained with cross-entropy on token pairs (Eq. 14), assuming the standard sequence-modeling objective transfers to semantic-detail pairs.
invented entities (1)
  • Semantic-Detail Associated Dual Codebook
    purpose: Decouple image patches into a compact semantic code and a residual detail code to ease autoregressive prediction.
    Evidence is internal to the paper (Tables 1, 7); no external falsifiable prediction outside this benchmark is offered.

how reviews work

0 comments
Cite this review

Pith. "Pith review of IAR2: Improving Autoregressive Visual Generation with Semantic-Detail Associated Token Prediction." pith.science (2026). https://pith.science/paper/IGZOZFD7

@misc{pith2026251006928,
  author       = {Pith},
  title        = {Pith review of: IAR2: Improving Autoregressive Visual Generation with Semantic-Detail Associated Token Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IGZOZFD7}},
  note         = {Machine review of arXiv:2510.06928}
}
read the original abstract

Autoregressive models have emerged as a powerful paradigm for visual content creation, but often overlook the intrinsic structural properties of visual data. Our prior work, IAR, initiated a direction to address this by reorganizing the visual codebook based on embedding similarity, thereby improving generation robustness. However, it is constrained by the rigidity of pre-trained codebooks and the inaccuracies of hard, uniform clustering. To overcome these limitations, we propose IAR2, an advanced autoregressive framework that enables a hierarchical semantic-detail synthesis process. At the core of IAR2 is a novel Semantic-Detail Associated Dual Codebook, which decouples image representations into a semantic codebook for global semantic information and a detail codebook for fine-grained refinements. It expands the quantization capacity from a linear to a polynomial scale, significantly enhancing expressiveness. To accommodate this dual representation, we propose a Semantic-Detail Autoregressive Prediction scheme coupled with a Local-Context Enhanced Autoregressive Head, which performs hierarchical prediction-first the semantic token, then the detail token-while leveraging a local context window to enhance spatial coherence. Furthermore, for conditional generation, we introduce a Progressive Attention-Guided Adaptive CFG mechanism that dynamically modulates the guidance scale for each token based on its relevance to the condition and its temporal position in the generation sequence, improving conditional alignment without sacrificing realism. Extensive experiments demonstrate that IAR2 sets a new state-of-the-art for autoregressive image generation, achieving a FID of 1.50 on ImageNet. Our model not only surpasses previous methods in performance but also demonstrates superior computational efficiency, highlighting the effectiveness of our structured, coarse-to-fine generation strategy.

Figures

Figures reproduced from arXiv: 2510.06928 by the authors.

Figure 1
Figure 1. Performance Comparison with the state-of-the-art [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The IAR Framework: IAR begins by rearranging its [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. (a) The MSE and LPIPS between the source image and the reconstructed image under different code distances. (b) Visualization of decoded images at varying code distances. Formally, given the codebook Z = {zi} N i=1, the goal is to find a mapping M(·) that minimizes the distance between consecutive embeddings: M = arg min M N X−1 i=1 ∥zM(i) , zM(i+1)∥. (1) This optimization can be reduced to Hamiltonian-path problem, … view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Impact of codebook size on reconstruction and [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: IAR2 consists of three main modules: 1) The [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Comparison between conventional fixed CFG (a) and [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Analysis on the Generation Hyperparameters: (a) CFG strength; (b) Parameter number; and (3) Training epoch. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: The training loss curves for the semantic cross [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: The generated images for balloon, house finch, triumphal arch, and breakwater by IAR2-XL. [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]
Figure 10
Figure 10. Figure 10: The generated images for alp, Arctic fox, marmot, and liner by IAR2-XL. [PITH_FULL_IMAGE:figures/full_fig_p021_10.png]
Figure 11
Figure 11. Figure 11: The generated images for coyote, schooner, stupa, and daisy by IAR2-XL. [PITH_FULL_IMAGE:figures/full_fig_p022_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 16 linked inside Pith

  1. [1]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P . Abbeel, “Denoising diffusion probabilistic models,” inNeurIPS, 2020. 1

    2020

  2. [2]

    Generative adversarial networks,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial networks,”Commun Acm, 2020. 1, 3

    2020

  3. [3]

    Autoregressive model beats diffusion: Llama for scalable image generation,

    P . Sun, Y. Jiang, S. Chen, S. Zhang, B. Peng, P . Luo, and Z. Yuan, “Autoregressive model beats diffusion: Llama for scalable image generation,”arXiv preprint arXiv:2406.06525, 2024. 1, 3, 4, 6, 8, 10, 11, 14, 18

    arXiv 2024

  4. [4]

    Visual autoregressive modeling: Scalable image generation via next-scale prediction,

    K. Tian, Y. Jiang, Z. Yuan, B. Peng, and L. Wang, “Visual autoregressive modeling: Scalable image generation via next-scale prediction,” inNeurIPS, 2024. 1, 3, 11

    2024

  5. [5]

    Maskgit: Masked generative image transformer,

    H. Chang, H. Zhang, L. Jiang, C. Liu, and W. T. Freeman, “Maskgit: Masked generative image transformer,” inCVPR, 2022. 1, 11, 18

    2022

  6. [6]

    Meissonic: Revitalizing masked generative transformers for efficient high-resolution text-to-image synthesis,

    J. Bai, T. Ye, W. Chow, E. Song, Q.-G. Chen, X. Li, Z. Dong, L. Zhu, and S. Yan, “Meissonic: Revitalizing masked generative transformers for efficient high-resolution text-to-image synthesis,” arXiv preprint arXiv:2410.08261, 2024. 1

    arXiv 2024

  7. [7]

    Language models are unsupervised multitask learners,

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskeveret al., “Language models are unsupervised multitask learners,”OpenAI blog, 2019. 1, 3

    2019

  8. [8]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, “Bert: Pre-training of deep bidirectional transformers for language understanding,”arXiv preprint arXiv:1810.04805, 2018. 1, 3

    arXiv 2018

Show all 57 references
  1. [9]

    Improving autoregressive visual generation with cluster- oriented token prediction,

    T. Hu, J. Zhang, R. Yi, J. Weng, Y. Wang, X. Zeng, Z. Xue, and L. Ma, “Improving autoregressive visual generation with cluster- oriented token prediction,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 9351–9360. 1, 6, 11

    2025

  2. [10]

    Neural discrete representation learning,

    A. Van Den Oord, O. Vinyalset al., “Neural discrete representation learning,”Advances in neural information processing systems, vol. 30,

  3. [11]

    Taming transformers for high-resolution image synthesis,

    P . Esser, R. Rombach, and B. Ommer, “Taming transformers for high-resolution image synthesis,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 12 873–12 883. 2, 3

    2021

  4. [12]

    Vector-quantized image modeling with improved vqgan,

    J. Yu, X. Li, J. Y. Koh, H. Zhang, R. Pang, J. Qin, A. Ku, Y. Xu, J. Baldridge, and Y. Wu, “Vector-quantized image modeling with improved vqgan,”arXiv preprint arXiv:2110.04627, 2021. 2

    2021 arXiv

  5. [13]

    Autoregressive image generation using residual quantization,

    D. Lee, C. Kim, S. Kim, M. Cho, and W.-S. Han, “Autoregressive image generation using residual quantization,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 11 523–11 532. 2

    2022

  6. [14]

    Factorized visual tokenization and generation,

    Z. Bai, J. Gao, Z. Gao, P . Wang, Z. Zhang, T. He, and M. Z. Shou, “Factorized visual tokenization and generation,”arXiv preprint arXiv:2411.16681, 2024. 3

    2024 arXiv

  7. [15]

    Unitok: A unified tokenizer for visual generation and understanding,

    C. Ma, Y. Jiang, J. Wu, J. Yang, X. Yu, Z. Yuan, B. Peng, and X. Qi, “Unitok: A unified tokenizer for visual generation and understanding,”arXiv preprint arXiv:2502.20321, 2025. 3

    2025

  8. [16]

    Tokenflow: Unified image tokenizer for multimodal understanding and generation,

    L. Qu, H. Zhang, Y. Liu, X. Wang, Y. Jiang, Y. Gao, H. Ye, D. K. Du, Z. Yuan, and X. Wu, “Tokenflow: Unified image tokenizer for multimodal understanding and generation,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 2545–

    2025

  9. [17]

    Magvit: Masked generative video transformer,

    L. Yu, Y. Cheng, K. Sohn, J. Lezama, H. Zhang, H. Chang, A. G. Hauptmann, M.-H. Yang, Y. Hao, I. Essaet al., “Magvit: Masked generative video transformer,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 10 459–10 469. 3

    2023

  10. [18]

    Open- magvit2: An open-source project toward democratizing auto- regressive visual generation,

    Z. Luo, F. Shi, Y. Ge, Y. Yang, L. Wang, and Y. Shan, “Open- magvit2: An open-source project toward democratizing auto- regressive visual generation,”arXiv preprint arXiv:2409.04410,

    arXiv

  11. [19]

    Dualtoken: Towards unifying visual understanding and generation with dual visual vocabularies,

    W. Song, Y. Wang, Z. Song, Y. Li, H. Sun, W. Chen, Z. Zhou, J. Xu, J. Wang, and K. Yu, “Dualtoken: Towards unifying visual understanding and generation with dual visual vocabularies,” arXiv preprint arXiv:2503.14324, 2025. 3, 14, 15

    2025 arXiv

  12. [20]

    Conditional generative adversarial nets,

    M. Mirza and S. Osindero, “Conditional generative adversarial nets,”arXiv preprint arXiv:1411.1784, 2014. 3

    2014 arXiv

  13. [21]

    Unsupervised represen- tation learning with deep convolutional generative adversarial networks,

    A. Radford, L. Metz, and S. Chintala, “Unsupervised represen- tation learning with deep convolutional generative adversarial networks,”arXiv preprint arXiv:1511.06434, 2015. 3

    2015 arXiv

  14. [22]

    Improved techniques for training gans,

    T. Salimans, I. Goodfellow, W. Zaremba, V . Cheung, A. Radford, and X. Chen, “Improved techniques for training gans,”Advances in neural information processing systems, vol. 29, 2016. 3

    2016

  15. [23]

    Image-to-image translation with conditional adversarial networks,

    P . Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros, “Image-to-image translation with conditional adversarial networks,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 1125–1134. 3

    2017

  16. [24]

    Unpaired image-to- image translation using cycle-consistent adversarial networks,

    J.-Y. Zhu, T. Park, P . Isola, and A. A. Efros, “Unpaired image-to- image translation using cycle-consistent adversarial networks,” in Proceedings of the IEEE international conference on computer vision, 2017, pp. 2223–2232. 3

    2017

  17. [25]

    A style-based generator archi- tecture for generative adversarial networks,

    T. Karras, S. Laine, and T. Aila, “A style-based generator archi- tecture for generative adversarial networks,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 4401–4410. 3

    2019

  18. [26]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P . Abbeel, “Denoising diffusion probabilistic models,”Advances in neural information processing systems, vol. 33, pp. 6840–6851, 2020. 3

    2020

  19. [27]

    Improved denoising diffusion probabilistic models,

    A. Q. Nichol and P . Dhariwal, “Improved denoising diffusion probabilistic models,” inInternational conference on machine learn- ing. PMLR, 2021, pp. 8162–8171. 3

    2021

  20. [28]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,”arXiv preprint arXiv:2010.02502, 2020. 3

    2010 arXiv

  21. [29]

    Diffusion models beat gans on image synthesis,

    P . Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,”Advances in neural information processing systems, vol. 34, pp. 8780–8794, 2021. 3

    2021

  22. [30]

    Photorealistic text-to-image diffusion models with deep language understanding,

    C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. L. Denton, K. Ghasemipour, R. Gontijo Lopes, B. Karagol Ayan, T. Salimans et al., “Photorealistic text-to-image diffusion models with deep language understanding,”Advances in neural information processing systems, vol. 35, pp...

    2022

  23. [31]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P . Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 684–10 695. 3

    2022

  24. [32]

    Gpt-4 technical report,

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkatet al., “Gpt-4 technical report,”arXiv preprint arXiv:2303.08774, 2023. 3

    2023 arXiv

  25. [33]

    Zero-shot text-to-image generation,

    A. Ramesh, M. Pavlov, G. Goh, S. Gray, C. Voss, A. Radford, M. Chen, and I. Sutskever, “Zero-shot text-to-image generation,” inInternational conference on machine learning. Pmlr, 2021, pp. 8821–

    2021

  26. [34]

    Scaling autoregressive models for content-rich text-to-image generation,

    J. Yu, Y. Xu, J. Y. Koh, T. Luong, G. Baid, Z. Wang, V . Vasudevan, A. Ku, Y. Yang, B. K. Ayanet al., “Scaling autoregressive models for content-rich text-to-image generation,”arXiv preprint arXiv:2206.10789, vol. 2, no. 3, p. 5, 2022. 3

    2022 arXiv

  27. [35]

    Llama: Open and efficient foundation language models,

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozi`ere, N. Goyal, E. Hambro, F. Azharet al., “Llama: Open and efficient foundation language models,”arXiv preprint arXiv:2302.13971, 2023. 3

    2023 arXiv

  28. [36]

    Maskgit: Masked generative image transformer,

    H. Chang, H. Zhang, L. Jiang, C. Liu, and W. T. Freeman, “Maskgit: Masked generative image transformer,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 11 315–11 325. 3

    2022

  29. [37]

    Muse: Text-to-image generation via masked generative trans- formers,

    H. Chang, H. Zhang, J. Barber, A. Maschinot, J. Lezama, L. Jiang, M.-H. Yang, K. Murphy, W. T. Freeman, M. Rubinsteinet al., “Muse: Text-to-image generation via masked generative trans- formers,”arXiv preprint arXiv:2301.00704, 2023. 3

    2023 arXiv

  30. [38]

    Language model beats diffusion–tokenizer is key to visual generation,

    L. Yu, J. Lezama, N. B. Gundavarapu, L. Versari, K. Sohn, D. Min- nen, Y. Cheng, V . Birodkar, A. Gupta, X. Guet al., “Language model beats diffusion–tokenizer is key to visual generation,”arXiv preprint arXiv:2310.05737, 2023. 3

    2023 arXiv

  31. [39]

    Taming transformers for high-resolution image synthesis,

    P . Esser, R. Rombach, and B. Ommer, “Taming transformers for high-resolution image synthesis,” inCVPR, 2021. 4, 6, 7, 11, 18

    2021

  32. [40]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P . Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” inCVPR, 2018. 4

    2018

  33. [41]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” inCVPR,

  34. [42]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochre- iter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,” inNeurIPS, 2017. 10

    2017

  35. [43]

    Improved techniques for training gans,

    T. Salimans, I. Goodfellow, W. Zaremba, V . Cheung, A. Radford, and X. Chen, “Improved techniques for training gans,” inNeurIPS,

  36. [44]

    Improved precision and recall metric for assessing generative models,

    T. Kynk ¨a¨anniemi, T. Karras, S. Laine, J. Lehtinen, and T. Aila, “Improved precision and recall metric for assessing generative models,” inNeurIPS, 2019. 11

    2019

  37. [45]

    Large scale gan training for high fidelity natural image synthesis,

    A. Brock, J. Donahue, and K. Simonyan, “Large scale gan training for high fidelity natural image synthesis,”arXiv preprint arXiv:1809.11096, 2018. 11

    2018 arXiv

  38. [46]

    Scaling up gans for text-to-image synthesis,

    M. Kang, J.-Y. Zhu, R. Zhang, J. Park, E. Shechtman, S. Paris, and T. Park, “Scaling up gans for text-to-image synthesis,” inCVPR,

  39. [47]

    Stylegan-xl: Scaling stylegan to large diverse datasets,

    A. Sauer, K. Schwarz, and A. Geiger, “Stylegan-xl: Scaling stylegan to large diverse datasets,” inACM SIGGRAPH, 2022. 11

    2022

  40. [48]

    Diffusion models beat gans on image synthesis,

    P . Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,” inNeurIPS, 2021. 11

    2021

  41. [49]

    Cascaded diffusion models for high fidelity image generation,

    J. Ho, C. Saharia, W. Chan, D. J. Fleet, M. Norouzi, and T. Salimans, “Cascaded diffusion models for high fidelity image generation,” JMLR, 2022. 11

    2022

  42. [50]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P . Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” inCVPR, 2022. 11

    2022

  43. [51]

    Scalable diffusion models with transform- ers,

    W. Peebles and S. Xie, “Scalable diffusion models with transform- ers,” inICCV, 2023. 11

    2023

  44. [52]

    Vector-quantized image modeling with improved vqgan,

    J. Yu, X. Li, J. Y. Koh, H. Zhang, R. Pang, J. Qin, A. Ku, Y. Xu, J. Baldridge, and Y. Wu, “Vector-quantized image modeling with improved vqgan,” inICLR, 2021. 11

    2021

  45. [53]

    Autoregressive image generation using residual quantization,

    D. Lee, C. Kim, S. Kim, M. Cho, and W.-S. Han, “Autoregressive image generation using residual quantization,” inCVPR, 2022. 11

    2022

  46. [54]

    Min- ing top-k frequent itemsets through progressive sampling,

    A. Pietracaprina, M. Riondato, E. Upfal, and F. Vandin, “Min- ing top-k frequent itemsets through progressive sampling,” DATAMINE, 2010. 18

    2010

  47. [55]

    The curious case of neural text degeneration,

    A. Holtzman, J. Buys, L. Du, M. Forbes, and Y. Choi, “The curious case of neural text degeneration,” inICLR, 2019. 18

    2019

  48. [56]

    A learning algorithm for boltzmann machines,

    D. H. Ackley, G. E. Hinton, and T. J. Sejnowski, “A learning algorithm for boltzmann machines,”Cognitive Science, 1985. 18

    1985

  49. [57]

    Synthetic literature: Writing science fiction in a co-creative process,

    E. Manjavacas, F. Karsdorp, B. Burtenshaw, and M. Kestemont, “Synthetic literature: Writing science fiction in a co-creative process,” inCCNLG, 2017. 18 18 Appendix AOVERVIEW In this supplementary material, more details about the proposed IAR2 method and more experimental resu...

    2017

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.